Compression : creative or technical?

Hello everyone,

a couple of interesting points has been made elsewhere and we thought it we would be good for the VWG to do a summary on ACESCentral.

@Troy_James_Sobotka started the thread this way :

I am unsure that the volume compression is even remotely as “creative subjective” as some would believe; it’s a finite analytical / technical compression of the amplitude that should maintain the tonality distribution.

To which @nick replied :

Isn’t a compression always subjective? Where the compression curve begins and how far out gets compressed to the boundary are choices, and there are no “correct” answers.

Followed by some interesting examples, provided by Troy :

Whether or not aforementioned aesthetic compression function is acceptable is not as wide open creative either; there are finite boundaries that even casual observers will deem unacceptable. Over compression is a very real problem at 8 bit consumption levels, and yields rather garish imagery as demonstrated.

But the primary point I would make is that if we concretely identify what “tone” is, and I’d hope most here would agree more or less on a definition, there is precisely one correct compression target.

The curve has a direct relation to effected luminance. But it’s actually affecting radiometric-like energy. Feeding an achromatic sweep pattern yields one luminance series.

That may or may not be acceptable; too much compression will flatten tonality. Again, this is not an openly creative / subjective thing. I mean I guess it is, if one accepts “Wow that looks woeful” as potentially a desired target.

When we feed a chromatic series to the curve, the input energy is the mechanic, and the sensation is luminance. The affected value is energy-like, the effected value is luminance.

Therefore we are able to take any chromatic series and, subject to the nature of the signal compression, derive the luminance. And for any given curve, there is again, exactly one luminance series.

Example with an asymptote curve; the values “s” curve to an asymptote. This means that all energy-like values in the triplet have an input of open radiometric-like domain X to output closed domain display / optically linear Y. Here’s what happens…


We can already see that we have a huge problem here. The inverse falloff of her cheek should be far greater. Yet all S curves will do this if we follow the general principles of energy transport.

Here, as X approaches infinity, Y approach 100% optically linear. Hence that tonality collapse.

Now let’s see what happens if we loosen the asymptote and let the curve “shoot upwards” beyond the limited domain we specify with a constant slope. This is via Colour’s awesome little extrapolate tool I might add, so thanks to Thomas (yet again)…


Note how we are now recovering some of the inverse square law of the energy because we are now not plateauing, and instead letting luminance “carry on” up the optically linear domain output. Also note that we still have problems. I’d suspect most folks would agree that the inverse square off of her cheek is not yet appropriate of the luminance we would expect on perceptual ingestion of the chromatic scene. Also note that there are heavy “cusps” where the curve is still not ascending enough to reveal tonality. This is the tonality posterization / collapse effect here.

Finally, let’s see if we loosen the upper end of the S curve such that the angle carries on with a more “reasonable” slope.


Presto. Sure we can all bicker about which is more acceptable here, but I’d be rather confident that the latter is closer to the estimated apparent luminance than the first two if we canvassed most folks here.

So what can we glean from this? The answer is that again, for any given aesthetic compression of the signal amplitude, there is precisely one luminance series. That’s critical. It also permits further inferences.

A few observations:

  1. That the notion of a perfectly analogous S curve as per the Hurter Driffield density curves, under variable chromaticities due to film dye density in juxtaposition to the notion of constant chromaticity digital RGB emissions is problematic in a fundamental sense.
  2. That we can clearly identify posterization based off of the feedback from a luminance calculation for any given signal compression curve long before we add in chroma.
  3. That based on the difference between the idealized luminance feedback for the compression curve, we can simultaneously deduce the optimal volume compression using a chromaticity linear mechanic; it’s a minimization problem essentially where we traverse the chromaticity linear lines to push back luminance to the appropriate levels where the displays are incapable of producing fully chromatic variants.

Feel free to join the conversation !


Doesn’t that assume that the optimal volume compression is based on staying chromaticity linear? I’m not sure that is a given, since straight lines in CIExy are not lines of perceptual hue. It seems to me that the deduction that there is only one “correct” solution is based on taking some things as fixed which in fact may be variable.

8 bit is quality SDR. Mainstream SDR is 6 bit + FRC so it is a really good point.

CIExy might not have straight lines of perceptual hue but I’d have a look at repeating the experiment in ICtCp which uses a LMS transform to compute I. My feeling is that the results will carry over.

Exactly. There are various possibilities available. That is why I am not convinced of the one unique solution.

I approach this by working inversely. If a “tone” curve represents tonality, it tells us that it is bound in large to luminance. As a result, luminance dictates that our ratios must be maintained relative to the mixtures at hand.

Perceptual is a negotiation between two different domains. This is a tangential issue in my mind, as the majority of “perceptual” things happen as a byproduct of presenting optically linear energy to the systems. That is, we can indeed use perceptual facets to assert that the compression holds a perceptual hue linearity, but that is a later pass between chromaticity linear domains of the open radiometric model to our compressed signal in the closed domain. No need to muddle domains.

I would add that it is absolutely critical to keep both the open and closed compressed signal domain chromaticity linear. Failing to do so will explode all existing DCC tooling that lean on RGB values, which in turn are chromaticity linear. Think about keying, despill, pulling a secondary in a grading tool, etc. Hence why pushing the perceptual “negotiation” further down the chain is important.

1 Like

I guess @Troy_James_Sobotka tries to say that perception is happening anyway when we “see”. So a display medium shall not include perceptual modelling.

But what if the display cannot show the full signal range?
Would it not be benifical to compress the signal into the display range in a similar way we would have compress the original signal in the first place?

And I really think there is a fundamental perceptual difference beyween looking into a scene vs looking at a reproduction of a scene.


Could indeed be desirable! For example, we might want a “perceptually hue linear” chromaticity compression of the amplitude. That however could easily be negotiated between the difference of our two signals, to keep the two signals “rendering-like” agnostic, and only add in those sorts of flourishes in as a modification to introduce a third state.

Also, as touched upon below, the idea of tonality somewhat transcends the creative choice of a perceptual modification. It is worth considering tonality and signal compression as a separate component here, perhaps.

Couldn’t agree more!

With that said, it’s image making, and there are plenty of contexts where we aren’t expecting a “as though we are standing there” image, but rather an image formation convention? Film sure wasn’t “as though we were standing there” medium, and it was a decent creative medium for a century. Nor was it “perceptual”, but more chemical-mechanical.

I’d hope that it would be feasible to take the generalized concepts around “tonality” and at least get to a firm foundation for deducing a potential signal compression ground truth, without becoming a rigid dive into the creatively optional aspects of perceptual facets.



I got the permission by Rory Gordon to quote her :

  • What is tonality ?
  • The handling of gray scale information in a patterned, predictable way.